Machine learning, federated learning, and bioinformatics methods for biomedical and genomic data in onco-haematological diseases and neurodegenerative disorders

Biomedicine is currently facing a paradoxical situation: despite unprecedented health data volumes from several sources of information, data analysis, privacy compliance, and equitable access to healthcare pose critical challenges. This thesis bridges AI development with clinical research in oncohaematology and neuroscience to address these needs, following a multidisciplinary approach. Structured in two synergistic parts, it reflects the research and professional journey of the author. The first section is devoted to the investigation on Federated Learning, a privacy-preserving approach to perform distributed training of Artificial Intelligence models. It presents the validation of the Federated Learning platform of the EU project GenoMed4All, in a real-world onco-haematological use case, demonstrating that collaborative training outperforms isolated training. It further details the implementation of the real data management workflow for the SYNTHEMA Federated Learning infrastructure. Then, it finally addresses the challenge of heterogeneous and biased data by applying Federated Synthetic Data Sharing as an effective mitigation strategy. The second part focuses on bioinformatics and Machine Learning methods applied to neurological disorders. It presents the development of a diagnostic bioinformatic workflow for nanopore sequencing data, specialized for the characterization of patients affected by repeat expansion disorders. Then, an exploratory study on a cohort of patients affected by Myotonic Dystrophy type 1 aimed to deepen the genotype-phenotype correlation, by integrating multi-modal data, such as clinical, genomic, neuropsychological, and imaging. Finally, this thesis concludes with a study on mitochondrial diseases, applying Machine Learning approaches and comparing short- and long-read sequencing technologies, to characterize the mutational profiles of patients. In conclusion, this work offers interdisciplinary tools for biomedical challenges. It strongly advises Federated Learning not only as a privacy-preserving approach, but also to democratize healthcare and integrate under-resourced institutions into global advancements. It also provides frameworks for neurological disease analysis, specifically targeting precision medicine applications in biomedicine.